Update app.py
Browse files
app.py
CHANGED
@@ -3,90 +3,80 @@ import numpy as np
|
|
3 |
import h5py
|
4 |
import faiss
|
5 |
import json
|
6 |
-
from transformers import AutoTokenizer, AutoModel, AutoModelForMaskedLM
|
7 |
-
from sklearn.feature_extraction.text import TfidfVectorizer
|
8 |
-
from sklearn.metrics.pairwise import cosine_similarity
|
9 |
import re
|
10 |
from collections import Counter
|
11 |
import torch
|
12 |
from nltk.corpus import stopwords
|
13 |
from nltk.tokenize import word_tokenize
|
14 |
import nltk
|
|
|
|
|
|
|
15 |
|
16 |
# Download necessary NLTK data
|
17 |
nltk.download('stopwords', quiet=True)
|
18 |
nltk.download('punkt', quiet=True)
|
19 |
|
20 |
-
# Load
|
21 |
-
|
22 |
-
bert_lemma_tokenizer = AutoTokenizer.from_pretrained(bert_lemma_model_name)
|
23 |
-
bert_lemma_model = AutoModelForMaskedLM.from_pretrained(bert_lemma_model_name).to(torch.device("cuda" if torch.cuda.is_available() else "cpu"))
|
24 |
-
|
25 |
-
# Load BERT model for encoding search queries
|
26 |
-
bert_encode_model_name = 'anferico/bert-for-patents'
|
27 |
-
bert_encode_tokenizer = AutoTokenizer.from_pretrained(bert_encode_model_name)
|
28 |
-
bert_encode_model = AutoModel.from_pretrained(bert_encode_model_name)
|
29 |
-
|
30 |
-
def bert_lemmatize(text):
|
31 |
-
tokens = bert_lemma_tokenizer.tokenize(text)
|
32 |
-
input_ids = bert_lemma_tokenizer.convert_tokens_to_ids(tokens)
|
33 |
-
input_tensor = torch.tensor([input_ids]).to(torch.device("cuda" if torch.cuda.is_available() else "cpu"))
|
34 |
-
with torch.no_grad():
|
35 |
-
outputs = bert_lemma_model(input_tensor)
|
36 |
-
predictions = outputs.logits.argmax(dim=-1)
|
37 |
-
lemmatized_tokens = bert_lemma_tokenizer.convert_ids_to_tokens(predictions[0])
|
38 |
-
return ' '.join([token for token in lemmatized_tokens if token not in ['[CLS]', '[SEP]', '[PAD]']])
|
39 |
|
40 |
def preprocess_query(text):
|
41 |
-
#
|
42 |
-
text =
|
|
|
43 |
|
44 |
-
#
|
45 |
-
|
|
|
46 |
|
47 |
-
# Remove special characters
|
48 |
-
text = re.sub(r'[
|
|
|
|
|
|
|
|
|
49 |
|
50 |
# Tokenize
|
51 |
tokens = word_tokenize(text)
|
52 |
|
53 |
-
# Remove stopwords
|
54 |
stop_words = set(stopwords.words('english'))
|
55 |
-
tokens = [word for word in tokens if word not in stop_words]
|
|
|
|
|
|
|
56 |
|
57 |
-
#
|
58 |
-
|
59 |
|
60 |
-
#
|
61 |
-
|
|
|
62 |
|
63 |
-
|
|
|
|
|
|
|
64 |
|
65 |
def extract_key_features(text):
|
66 |
# For queries, we'll just preprocess and return all non-stopword terms
|
67 |
processed_text = preprocess_query(text)
|
68 |
-
|
69 |
# Split the processed text into individual terms
|
70 |
features = processed_text.split()
|
71 |
-
|
72 |
# Remove duplicates while preserving order
|
73 |
features = list(dict.fromkeys(features))
|
74 |
-
|
75 |
return features
|
76 |
|
77 |
-
def encode_texts(texts
|
78 |
-
|
79 |
-
|
80 |
-
outputs = bert_encode_model(**inputs)
|
81 |
-
embeddings = outputs.last_hidden_state.mean(dim=1)
|
82 |
-
return embeddings.numpy()
|
83 |
|
84 |
def load_data():
|
85 |
try:
|
86 |
with h5py.File('patent_embeddings.h5', 'r') as f:
|
87 |
embeddings = f['embeddings'][:]
|
88 |
patent_numbers = f['patent_numbers'][:]
|
89 |
-
|
90 |
metadata = {}
|
91 |
texts = []
|
92 |
with open('patent_metadata.jsonl', 'r') as f:
|
@@ -94,17 +84,13 @@ def load_data():
|
|
94 |
data = json.loads(line)
|
95 |
metadata[data['patent_number']] = data
|
96 |
texts.append(data['text'])
|
97 |
-
|
98 |
print(f"Embedding shape: {embeddings.shape}")
|
99 |
print(f"Number of patent numbers: {len(patent_numbers)}")
|
100 |
print(f"Number of metadata entries: {len(metadata)}")
|
101 |
-
|
102 |
return embeddings, patent_numbers, metadata, texts
|
103 |
-
except FileNotFoundError as e:
|
104 |
-
print(f"Error: Could not find file. {e}")
|
105 |
-
raise
|
106 |
except Exception as e:
|
107 |
-
print(f"An
|
108 |
raise
|
109 |
|
110 |
def compare_features(query_features, patent_features):
|
@@ -114,22 +100,21 @@ def compare_features(query_features, patent_features):
|
|
114 |
|
115 |
def hybrid_search(query, top_k=5):
|
116 |
print(f"Original query: {query}")
|
117 |
-
|
118 |
processed_query = preprocess_query(query)
|
119 |
query_features = extract_key_features(processed_query)
|
120 |
-
|
121 |
-
# Encode the processed query using the
|
122 |
query_embedding = encode_texts([processed_query])[0]
|
123 |
query_embedding = query_embedding / np.linalg.norm(query_embedding)
|
124 |
-
|
125 |
# Perform semantic similarity search
|
126 |
semantic_distances, semantic_indices = index.search(np.array([query_embedding]).astype('float32'), top_k * 2)
|
127 |
-
|
128 |
# Perform TF-IDF based search
|
129 |
query_tfidf = tfidf_vectorizer.transform([processed_query])
|
130 |
tfidf_similarities = cosine_similarity(query_tfidf, tfidf_matrix).flatten()
|
131 |
tfidf_indices = tfidf_similarities.argsort()[-top_k * 2:][::-1]
|
132 |
-
|
133 |
# Combine and rank results
|
134 |
combined_results = {}
|
135 |
for i, idx in enumerate(semantic_indices[0]):
|
@@ -142,7 +127,7 @@ def hybrid_search(query, top_k=5):
|
|
142 |
'common_features': common_features,
|
143 |
'text': text
|
144 |
}
|
145 |
-
|
146 |
for idx in tfidf_indices:
|
147 |
patent_number = patent_numbers[idx].decode('utf-8')
|
148 |
if patent_number not in combined_results:
|
@@ -154,10 +139,9 @@ def hybrid_search(query, top_k=5):
|
|
154 |
'common_features': common_features,
|
155 |
'text': text
|
156 |
}
|
157 |
-
|
158 |
# Sort and get top results
|
159 |
top_results = sorted(combined_results.items(), key=lambda x: x[1]['score'], reverse=True)[:top_k]
|
160 |
-
|
161 |
results = []
|
162 |
for patent_number, data in top_results:
|
163 |
result = f"Patent Number: {patent_number}\n"
|
@@ -165,19 +149,12 @@ def hybrid_search(query, top_k=5):
|
|
165 |
result += f"Combined Score: {data['score']:.4f}\n"
|
166 |
result += f"Common Key Features: {', '.join(data['common_features'])}\n\n"
|
167 |
results.append(result)
|
168 |
-
|
169 |
return "\n".join(results)
|
170 |
|
171 |
# Load data and prepare the FAISS index
|
172 |
embeddings, patent_numbers, metadata, texts = load_data()
|
173 |
|
174 |
-
# Check if the embedding dimensions match
|
175 |
-
if embeddings.shape[1] != encode_texts(["test"]).shape[1]:
|
176 |
-
print("Embedding dimensions do not match. Rebuilding FAISS index.")
|
177 |
-
# Rebuild embeddings using the new model
|
178 |
-
embeddings = encode_texts(texts)
|
179 |
-
embeddings = embeddings / np.linalg.norm(embeddings, axis=1, keepdims=True)
|
180 |
-
|
181 |
# Normalize embeddings for cosine similarity
|
182 |
embeddings = embeddings / np.linalg.norm(embeddings, axis=1, keepdims=True)
|
183 |
|
@@ -189,7 +166,7 @@ index.add(embeddings)
|
|
189 |
tfidf_vectorizer = TfidfVectorizer(stop_words='english')
|
190 |
tfidf_matrix = tfidf_vectorizer.fit_transform(texts)
|
191 |
|
192 |
-
# Create Gradio interface
|
193 |
iface = gr.Interface(
|
194 |
fn=hybrid_search,
|
195 |
inputs=[
|
@@ -202,4 +179,4 @@ iface = gr.Interface(
|
|
202 |
)
|
203 |
|
204 |
if __name__ == "__main__":
|
205 |
-
iface.launch()
|
|
|
3 |
import h5py
|
4 |
import faiss
|
5 |
import json
|
|
|
|
|
|
|
6 |
import re
|
7 |
from collections import Counter
|
8 |
import torch
|
9 |
from nltk.corpus import stopwords
|
10 |
from nltk.tokenize import word_tokenize
|
11 |
import nltk
|
12 |
+
from sentence_transformers import SentenceTransformer
|
13 |
+
from sklearn.feature_extraction.text import TfidfVectorizer
|
14 |
+
from sklearn.metrics.pairwise import cosine_similarity
|
15 |
|
16 |
# Download necessary NLTK data
|
17 |
nltk.download('stopwords', quiet=True)
|
18 |
nltk.download('punkt', quiet=True)
|
19 |
|
20 |
+
# Load SentenceTransformer model
|
21 |
+
model = SentenceTransformer('anferico/bert-for-patents')
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
22 |
|
23 |
def preprocess_query(text):
|
24 |
+
# Remove "[EN]" label and claim numbers
|
25 |
+
text = re.sub(r'\[EN\]\s*', '', text)
|
26 |
+
text = re.sub(r'^\d+\.\s*', '', text, flags=re.MULTILINE)
|
27 |
|
28 |
+
# Convert to lowercase while preserving acronyms and units
|
29 |
+
words = text.split()
|
30 |
+
text = ' '.join(word if word.isupper() or re.match(r'^\d+(\.\d+)?[a-zA-Z]+$', word) else word.lower() for word in words)
|
31 |
|
32 |
+
# Remove special characters except hyphens and periods in numbers
|
33 |
+
text = re.sub(r'[^\w\s\-.]', ' ', text)
|
34 |
+
text = re.sub(r'(?<!\d)\.(?!\d)', ' ', text) # Remove periods not in numbers
|
35 |
+
|
36 |
+
# Normalize spaces
|
37 |
+
text = re.sub(r'\s+', ' ', text).strip()
|
38 |
|
39 |
# Tokenize
|
40 |
tokens = word_tokenize(text)
|
41 |
|
42 |
+
# Remove stopwords
|
43 |
stop_words = set(stopwords.words('english'))
|
44 |
+
tokens = [word for word in tokens if word.lower() not in stop_words]
|
45 |
+
|
46 |
+
# Join tokens back into text
|
47 |
+
text = ' '.join(tokens)
|
48 |
|
49 |
+
# Preserve numerical values with units
|
50 |
+
text = re.sub(r'(\d+(\.\d+)?)([a-zA-Z]+)', r'\1_\3', text)
|
51 |
|
52 |
+
# Handle ranges and measurements
|
53 |
+
text = re.sub(r'(\d+(\.\d+)?)(\s*to\s*)(\d+(\.\d+)?)(\s*[a-zA-Z]+)', r'\1_to_\4_\6', text)
|
54 |
+
text = re.sub(r'between\s*(\d+(\.\d+)?)(\s*and\s*)(\d+(\.\d+)?)\s*([a-zA-Z]+)', r'between_\1_and_\4_\5', text)
|
55 |
|
56 |
+
# Preserve chemical formulas
|
57 |
+
text = re.sub(r'\b([A-Z][a-z]?\d*)+\b', lambda m: m.group().replace(' ', ''), text)
|
58 |
+
|
59 |
+
return text
|
60 |
|
61 |
def extract_key_features(text):
|
62 |
# For queries, we'll just preprocess and return all non-stopword terms
|
63 |
processed_text = preprocess_query(text)
|
|
|
64 |
# Split the processed text into individual terms
|
65 |
features = processed_text.split()
|
|
|
66 |
# Remove duplicates while preserving order
|
67 |
features = list(dict.fromkeys(features))
|
|
|
68 |
return features
|
69 |
|
70 |
+
def encode_texts(texts):
|
71 |
+
embeddings = model.encode(texts, show_progress_bar=True)
|
72 |
+
return embeddings
|
|
|
|
|
|
|
73 |
|
74 |
def load_data():
|
75 |
try:
|
76 |
with h5py.File('patent_embeddings.h5', 'r') as f:
|
77 |
embeddings = f['embeddings'][:]
|
78 |
patent_numbers = f['patent_numbers'][:]
|
79 |
+
|
80 |
metadata = {}
|
81 |
texts = []
|
82 |
with open('patent_metadata.jsonl', 'r') as f:
|
|
|
84 |
data = json.loads(line)
|
85 |
metadata[data['patent_number']] = data
|
86 |
texts.append(data['text'])
|
87 |
+
|
88 |
print(f"Embedding shape: {embeddings.shape}")
|
89 |
print(f"Number of patent numbers: {len(patent_numbers)}")
|
90 |
print(f"Number of metadata entries: {len(metadata)}")
|
|
|
91 |
return embeddings, patent_numbers, metadata, texts
|
|
|
|
|
|
|
92 |
except Exception as e:
|
93 |
+
print(f"An error occurred while loading data: {e}")
|
94 |
raise
|
95 |
|
96 |
def compare_features(query_features, patent_features):
|
|
|
100 |
|
101 |
def hybrid_search(query, top_k=5):
|
102 |
print(f"Original query: {query}")
|
|
|
103 |
processed_query = preprocess_query(query)
|
104 |
query_features = extract_key_features(processed_query)
|
105 |
+
|
106 |
+
# Encode the processed query using the SentenceTransformer model
|
107 |
query_embedding = encode_texts([processed_query])[0]
|
108 |
query_embedding = query_embedding / np.linalg.norm(query_embedding)
|
109 |
+
|
110 |
# Perform semantic similarity search
|
111 |
semantic_distances, semantic_indices = index.search(np.array([query_embedding]).astype('float32'), top_k * 2)
|
112 |
+
|
113 |
# Perform TF-IDF based search
|
114 |
query_tfidf = tfidf_vectorizer.transform([processed_query])
|
115 |
tfidf_similarities = cosine_similarity(query_tfidf, tfidf_matrix).flatten()
|
116 |
tfidf_indices = tfidf_similarities.argsort()[-top_k * 2:][::-1]
|
117 |
+
|
118 |
# Combine and rank results
|
119 |
combined_results = {}
|
120 |
for i, idx in enumerate(semantic_indices[0]):
|
|
|
127 |
'common_features': common_features,
|
128 |
'text': text
|
129 |
}
|
130 |
+
|
131 |
for idx in tfidf_indices:
|
132 |
patent_number = patent_numbers[idx].decode('utf-8')
|
133 |
if patent_number not in combined_results:
|
|
|
139 |
'common_features': common_features,
|
140 |
'text': text
|
141 |
}
|
142 |
+
|
143 |
# Sort and get top results
|
144 |
top_results = sorted(combined_results.items(), key=lambda x: x[1]['score'], reverse=True)[:top_k]
|
|
|
145 |
results = []
|
146 |
for patent_number, data in top_results:
|
147 |
result = f"Patent Number: {patent_number}\n"
|
|
|
149 |
result += f"Combined Score: {data['score']:.4f}\n"
|
150 |
result += f"Common Key Features: {', '.join(data['common_features'])}\n\n"
|
151 |
results.append(result)
|
152 |
+
|
153 |
return "\n".join(results)
|
154 |
|
155 |
# Load data and prepare the FAISS index
|
156 |
embeddings, patent_numbers, metadata, texts = load_data()
|
157 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
158 |
# Normalize embeddings for cosine similarity
|
159 |
embeddings = embeddings / np.linalg.norm(embeddings, axis=1, keepdims=True)
|
160 |
|
|
|
166 |
tfidf_vectorizer = TfidfVectorizer(stop_words='english')
|
167 |
tfidf_matrix = tfidf_vectorizer.fit_transform(texts)
|
168 |
|
169 |
+
# Create Gradio interface
|
170 |
iface = gr.Interface(
|
171 |
fn=hybrid_search,
|
172 |
inputs=[
|
|
|
179 |
)
|
180 |
|
181 |
if __name__ == "__main__":
|
182 |
+
iface.launch()
|